ICE Scoring

Start with 10–20 candidate tests in a single table in Attio or Notion, then enrich each row in Clay with source, segment, channel, and expected motion. In 2026, a useful example is pulling 40 ideas from Common Room product signals, then using Clay Chat to summarise them into one-line hypotheses like “priority account users who hit feature X convert 18% better.” No spreadsheet theatre: one row, one hypothesis, one owner.

Give Impact a 1–10 score based on upside, not vanity. A 2026 AI-native example is using PostHog funnels plus Claude Code to estimate whether a homepage change could move trial-to-paid by 2% or 0.2%, then letting Clay write the score back into the record. If the test can affect pipeline, activation, or retained revenue, it gets a higher number; if it only changes click-through, it does not.

Confidence measures how much evidence you already have, and 2026 teams should automate most of it. A clean example is a Claude 4.5 Sonnet prompt reading PostHog session replays, Common Room intent, and 11x SDR notes, then returning a confidence score with a 2-line rationale. If 3 sources agree, confidence rises; if the idea comes from one loud stakeholder, it stays low. No evidence, no high score.

Ease is the inverse of engineering drag, so score the path to live, not the size of the idea. In 2026, use Framer + Relume for landing pages, Trigger.dev for workflow glue, and Vercel for fast deploys, then assign a higher Ease score when the experiment can ship in 1 day. A landing-page test with AdCreative.ai copy and a Framer variant is easier than a checkout rewrite in Next.js.

Multiply or average the 3 scores, then rank the backlog in Attio or Clay and only ship the top 3. A practical 2026 example is sending the winning ideas to 11x for outbound follow-up, using n8n to trigger the asset build, and logging results back into PostHog after 7 days. If an experiment scores high on Impact but low on Ease, it waits; if it is easy but tiny, it dies.